Regression with interaction term (SPSS)?
Hi everyone, I have a quick question. I would like to conduct a regression analysis with an interaction term. However, there are a few prerequisites for this that are a bit confusing for me. I'll leave out the two prerequisites that the variables must be metrically scaled and normally distributed (they already apply). Now I come to the prerequisite that there must be a linear relationship between the variables (no multicorrelation). I have already checked the independent variable (stress) and the dependent variable (drug use) and no significant correlation emerged. In the next step, however, I wanted to find out whether impulsivity moderates the relationship between stress and drug use. This brings me to my question. Does the regression analysis with an interaction term still make sense if neither stress and drug use correlate with each other/have a linear relationship, nor does impulsivity have a linear relationship with drug use? How do I proceed in the next step? Do I say that the calculation cannot be performed and the hypothesis is not confirmed, or do other steps need to be taken?
Thank you in advance 🙂
(2 votes)
Loading...
Your sample seems pretty small. In this case, “not significant” would not be particularly informative. Can be too low a statistical power that ensures false-negative results, or that there are no connections. You don’t know.
So my sample has a size of 100 people. There is no linear relationship between the variables. Do not even correlate the independent with the dependent variable.
Just calculate your planned moderation analysis and interpret the coefficients. “significant” bivariate correlations between the predictors and the dependent variable are not a prerequisite. The moderator should best not be correlated with the other UV or mi of the AV, and the effect of the variable actually of interest can just be deteriorated by a not taken into account moderator effect, see simplified example here:
(The average value of AV is just as high in the GESAMTEN Low TQM condition as in the GESAMEN High TQM condition, i.e. apparently no effect of TQM; only by adding CP as moderator you see what is actually going on).
However, I do not know what is meant by “nonlinear” whether a small correlation, or actually a U-shaped or J-shaped or otherwise nonlinear relationship.
I must confess, I do not make sense or purpose. I answered the questions I asked for the best of my knowledge, so I’ll stop at this point.
I added an image, there are the variables (lengthy, independent, moderator) included. So I used the Loess function to create the straight line if this was correct. And what you see in the picture came out.
It is not so much about distortion, but about correctly estimated standard errors.
Normal distribution of the variables is irrelevant. In any case, the regression model’s regressions should originate from a normal distributed population, but also from about n > 30 no matter.
“Linear relationship” means that the relationship should be linear, non-linear (e.g. exponential, J-shaped, U-shaped).
It is important to have homoskedascity.
I have already described and illustrated with an example that bivariate correlations between UV and AV are not a prerequisite.
If your faculty or supervisors make other requirements, it’s just that. I can only tell you what’s actually happening.
So if I don’t täusche you check the linear relationship between the dependent, independent and moderator variable before starting with the moderator analysis. So there is definitely no linear relationship with me. There is no correlation between the independent variable and the dependent one, i.e. no correlation. Does this mean that I should not conduct the analysis best, as the results will be “distorted” anyway, as there are no normal distribution and other conditions, or should I take measures against it? The only requirement that is given is that there is no multicorrelinearity between the independent and the moderator variable.